An advanced hybrid deep learning framework for structural health monitoring of cable-stayed bridge model using fiber optic sensors

This paper proposes an efficient hybrid framework for structural health monitoring (SHM) of a laboratory-scale cable-stayed bridge, utilizing fiber optic sensors (FOS), specifically Fiber Bragg Grating (FBG) sensors, in conjunction with deep learning (DL) and ensemble methods. A one-dimensional convolutional neural network (1DCNN) is employed to extract local spatial features from the sensor data, followed by a long short-term memory (LSTM) network to model temporal dependencies. However, as conventional DL models often rely on softmax classifiers, which can be suboptimal in handling fine-grained or imbalanced damage patterns. We integrate eXtreme Gradient Boosting (XGBoost) as a final classifier to enhance robustness and classification precision. To address data scarcity, advanced augmentation techniques such as reversing, noise injection, window slicing, temporal warping, and shifting are employed. The augmented data is then segmented to enrich the dataset. The framework is validated using experimental time series data acquired from an FBG-based sensing system on a cable-stayed bridge model under various simulated damage scenarios. Notably, the best results are achieved with the hybrid models: LSTM-XGBoost, 1DCNN-XGBoost, and the proposed 1DCNN-LSTM-XGBoost, reaching accuracies of 82.2 %, 89.8 %, and 91.5 %, respectively. The proposed model demonstrates significant potential in damage assessment, a critical research topic in the SHM field.